In the loop: how AI is changing the legal sector

I recently came across the Japanese woodworking tradition of 'sashimono', an art of joinery where furniture is built without nails or screws. Each component is crafted with exacting precision so it fits seamlessly with the next. The strength of the structure comes not from force, but from how well the pieces align.

On its own, it’s fascinating. But what really excites me are the possibilities it opens up for reimagining how we could approach AI in legal work.

I've become used to seeing lawyers either throw AI at the entire task or reject it completely. But the most valuable uses aren't about replacement or resistance. They’re about alignment. The fit between human and machine is what matters most.

Think of AI not as an end-to-end solution, but as a component. Something that integrates into the broader workflow, bringing value only when shaped with intention and placed at the right points.

Here are five reflection points I’ve found useful when designing those joins:

1️⃣ Where does structure come from?
AI can offer a quick structure - a draft, a flow, a set of bullet points - but your role is to shape it so that it reflects the brief, the client, and the context.

2️⃣ What’s being assumed?
Outputs often carry assumptions about tone, risk, or what counts as standard. Spotting and adjusting these is where judgment becomes critical.

3️⃣ When should you step in?
Pinpoint the moments where your input adds the most value, whether that’s framing the task, refining the tone, or reviewing the output, and focus your attention there.

4️⃣ Is the handoff clear?
Unclear boundaries between human and machine lead to weaker results. Define the edge: where the model stops and your expertise begins.

5️⃣ Does the result carry conviction?
The best work does more than sound fluent. It takes a position. If the output feels smooth but says little, it likely needs a human voice.

Done well, this creates a kind of sashimono for legal work: a system of clean, intentional joins. No brute force. No awkward patchwork. Just clear integration between human thinking and AI support.

Where are you finding the most meaningful joins between your work and AI?

Working with AI has taught me a lot. And one lesson I’ve been thinking about this past week is that the most valuable outputs aren’t always the most polished. Sometimes the draft is clunky, the summary too thin, the redline misses what really matters.

But instead of dismissing those outputs, I’ve started asking: what is this making me notice?

Often, it’s the gaps that surface something more interesting - a missed nuance, a hidden assumption, or a better way to frame the task entirely. In moments like these, the model opens a loop. But it’s your reasoning that needs to close it.

And that, I’ve found, is where the real value lives. Not in replacing legal thinking, but in reframing it. When you treat the AI as a sparring partner (not an answer machine) you start to see the value in a different way. The outputs help you spot gaps, surface assumptions, and ask sharper questions.

For those starting to use AI tools for legal reasoning, or trying to get more value from imperfect outputs, here are a few approaches that I find work well:
1️⃣ Zoom out: Look beyond the surface. What’s been left out? What does the model assume to be true that might not be?
2️⃣ Zoom in: Read the language closely. Does it reflect the client’s tone, risk appetite, or the subtleties of the deal?
3️⃣ Use the silence: If something isn’t mentioned, that might be the most important clue. Sometimes what the model omits is more interesting than what it includes.
4️⃣ Switch tools when needed: If your model can’t handle long inputs or nuanced logic, try something with a larger context window or a more reasoning-optimised architecture. If your work requires real-time web browsing, there are several tools that excel at this.
5️⃣ Reframe the task: If the output isn’t hitting the mark, it might not be the model’s fault. It could be that the task wasn’t scoped quite right in the first place.

In my experience, the best AI outputs don’t always have to give you answers. They give you momentum. They help you move faster not by doing the work for you, but by reshaping how you approach the problem in the first place.

And that can often more valuable than getting something “right” the first time.

Where has AI made you stop and rethink, not just accelerate?

I’ve been working with AI-assisted gap analysis for a while now. On the surface, it’s a simple enough concept: compare Document A to Document B and spot what’s missing. It’s a very popular use case for AI to demonstrate value in legal work.

But the more I support lawyers with this exercise, the more I’m realising that gap analysis isn’t just a document comparison task. It’s a thinking exercise. Because AI can flag missing clauses, definitions, or concepts. It can point to differences in wording or structure. But it can’t tell you whether the missing thing matters. Whether it was deliberate. Whether it changes the risk profile. Whether it fits the client or the deal or the situation.

That’s still human work. That’s still judgment.

And that’s where AI tools become interesting. They don’t just help you work faster. They help you work sharper. Because when the mechanics of comparison are handled, your focus shifts to something deeper: interpreting the meaning behind the difference.

If you’re new to using AI for gap analysis, or just want to do it more effectively, you might find it helpful to consider the following questions:
1️⃣ Are you using the right tool for the job? A powerful AI tool with a large context window can make or break your output.
2️⃣ Have you structured the task clearly? Define exactly what kind of gaps you’re trying to surface.
3️⃣ What's the benchmark? Your precedent matters, and so does understanding its assumptions.
4️⃣ Who's doing the review? AI can point you to differences, but only legal judgment can assess their impact.
5️⃣ Is your workflow designed for iteration? You’ll often need to refine prompts or revisit outputs.

These are the kinds of questions I’m seeing come up more and more. Not just “what’s different?” but “why does the difference matter?”

And I think that’s where AI is starting to reshape legal thinking - not by replacing judgment, but by revealing where it’s most needed. I've touched on this idea in a previous post.

Where have you found your own thinking sharpened by a tool that simply showed you the obvious?

I can't stop thinking about the role of legal instinct in a world where AI can draft, redline, and review documents in seconds.

A clause might be perfectly structured. A summary might be factually sound. A gap analysis might tick every category. But none of that amounts to a decision. It doesn’t tell you when to challenge, when to accept, or when to reframe the issue altogether. That’s still human work. That’s still judgment.

By legal instinct, I don’t mean guesswork. I mean the muscle memory that comes from pattern recognition - the ability to spot when something feels off, even if the language is right. The skill to balance risk, commerciality, and context without needing a playbook. It’s that internal compass lawyers develop over time. And it doesn’t show up in track changes.

Now that AI is more embedded in legal workflows, I think that instinct is becoming more visible, not less. What used to be buried in the slow mechanics of drafting seems to be concentrated in sharper moments of input and review. Lawyers aren’t just producing a document anymore. They’re shaping the direction of it before and after the machine gets involved.

From what I’m seeing and hearing, legal instinct is showing up in three key places:

🔷 Before the prompt – framing the issue, scoping the task, and deciding what really matters
🔷 After the output – assessing whether the result reflects the client, the context, or the risk
🔷 Around the task – knowing when not to use AI at all

This shift isn’t theoretical. I’ve seen it across multiple teams. The best lawyers aren’t thinking less because of AI. They’re thinking more sharply because the routine work is no longer the bottleneck. And it’s making me question whether the document was ever the core deliverable, or if the value was always in the judgment that shaped it.

It also makes me wonder how legal education will adapt. When AI can do the writing, the real training ground might become the moments before and after the text appears. How do you teach that? How do you spot it? And how do future lawyers decide how much is enough, when overworking a deliverable might be adding less value, not more? Credit to the ever-impressive Minesh Tanna for making me think harder about this yesterday.

Has AI made your own instincts more visible? If so, where?

Human–AI interaction in legal decision-making has been on my mind for a while. Long before most legal teams were using AI the way they are today. From the beginning, I never saw AI replacing legal thinking. It always felt like augmentation. A tool to support, not override.

But lately, my view's started to shift.

The more I work with GenAI, and especially with agentic AI, the more I’m noticing something different. It’s not just helping lawyers think. It’s moving where that thinking happens.

In case you missed it, Simmons & Simmons recently partnered with Flank to explore agentic AI and begin building real-world legal agents. Working with these systems makes the shift impossible to ignore. This isn’t only about speed. It’s about 𝘸𝘩𝘦𝘳𝘦 thinking now fits into the process.

You notice it when a full draft lands after a single prompt. You’re not starting from scratch. You’re starting at review. And suddenly the focus isn’t writing. It’s scanning for whether good judgment is still in the mix.

I was chatting with the excellent Rob Cant about this last week. One point we kept returning to was how much depends on where you integrate the thinking. Legal thinking now has to be placed more intentionally in the workflow. Before AI, thinking was layered throughout the drafting process. Research, analysis, writing, reviewing, adjusting. You worked through the problem as you went.

Now, I believe AI compresses that middle.

The draft arrives structured and fluent. But the thinking still needs to happen. Just in different places. For instance:
🔷Before the prompt: framing the task, setting scope, deciding what matters
🔷 After the output: reviewing, refining, applying judgment
🔷 Around the task: deciding whether AI is the right tool at all

Drafting may no longer be the stage where the thinking unfolds. It may end up becoming the stage where the thinking gets applied.

That shift carries risk. Without intention upfront or proper review afterwards, you can end up approving something that sounds right but isn’t. The model can write. But it doesn’t understand client nuance, commercial context, or firm standards. That is still human work.

AI isn’t removing legal expertise. It’s revealing where that expertise needs to show up.

Where have you found your own thinking shifting as you work with AI?

I was reading BOND's recent report on the acceleration of AI adoption. The numbers really do speak for themselves. Adoption curves are steeper than any tech cycle we’ve seen. Usage is exploding. Costs are dropping. Model capabilities are improving almost weekly.

But while the models are moving fast, I think legal work hits a very different kind of friction. Not model size. Not token limits. But context.

To my mind, legal work has regularly depended on context that was rarely written down:
🔷 The way a particular client approaches risk
🔷 The things a partner expects to see, even if they never say it
🔷 The cultural and relational sensitivities that shape how advice is framed

When humans do legal work, that context travels silently. I believe lawyers carry it instinctively. It shapes how they draft, review, escalate, soften or sharpen a point. Legal professionals rarely stop to consciously define it. But AI needs all of that context spelled out. It can’t work with what was left unsaid.

This is where legal prompting quietly becomes something different. It isn’t just about giving the model instructions. It’s about surfacing the context you’ve never had to articulate before.

Here’s where I think that shows up most:
1️⃣ Audience and purpose - who is this for?
2️⃣ Tone - how formal or flexible should this sound?
3️⃣ Nuance - where should judgment override neutrality?
4️⃣ Positioning - what is commercially acceptable for this client or this deal?
5️⃣ Style and voice - how do we make this sound like us, not like generic legal output?

The more you supply, the better the model performs. The gap isn’t usually about model capability. It’s about the absence of information the model was never given. In that sense, AI adoption in legal isn’t just about learning how to use the tools. It’s about learning how to make the invisible parts of legal work visible.

That’s where the real skill is starting to emerge.

Where have you found AI forcing you to explain things you usually take for granted?

If you’re curious, you can access the BOND report here: https://lnkd.in/e47-Q_26

I’ve been doing some AI governance work recently, alongside the amazing Jack Mitchell and Peter Lee. The phrase ‘human in the loop’ keeps coming up. Over and over. It’s everywhere. And it makes sense.

But the other day, I had a random thought: What if we’ve got it the wrong way round? What if the AI is the one in the loop? Is that even a thing?

I did a bit of digging, and as it turns out, AI in the loop does exist as a concept. Not in legal circles, really. But it pops up in robotics, human-AI collaboration, design research. And the more I thought about it, the more it made sense.

Because when I look at how AI is being used in legal workflows, it doesn’t feel like the lawyer is supervising a system. It feels like the lawyer is designing the system. And the AI is sitting inside it. There’s a human shaping the task. A human prompting, reviewing, rejecting, adjusting. A human deciding where AI is allowed in, and where it isn’t.

In other words, the human is the workflow. The loop is already theirs. The AI is just… in it.

So, what does it look like when AI is in the loop?
🔹 When a lawyer sets the structure, and AI fills in the gaps
🔹 When a prompt is refined across iterations, not approved at the end
🔹 When AI contributes to the work, but never controls the direction
🔹 When human logic frames the task and human judgment wraps it up

That’s a different framing. Because if you see the AI as in the loop, not running it, you design things differently. In governance terms, it shifts the focus. It’s not just about who signs off at the end. It’s about who shapes the system from the start.

Anyway, just a thought. But one I keep coming back to.

Is AI in the loop a better way to describe how legal teams are actually using it?

Like everyone, I see considerable talk about the new skills lawyers need to thrive in an increasingly AI-driven world: Prompt engineering. Tool fluency. AI literacy. But I think there may be another side to this conversation. AI is highlighting something else, quietly, but powerfully.

It’s revealing the value of skills that have always mattered, but haven’t always been recognised. Skills that often lived in the margins of legal work - in edits, feedback loops, drafts that never made it to the client.

Now they’re front and centre.

Here are a few I've noticed resurface:
🔹 Clarity of thought - the ability to break down an issue with structure and intent
🔹 Sensitivity to tone - knowing when something is too flat, too robotic, too cautious
🔹 Attention to framing - catching when something is technically correct but strategically off
🔹 Confidence in judgment - not overediting or overexplaining what already works

These were never 'skills' in the traditional sense. I don’t think you’d see them on CVs or in pitches. But they shaped how good legal work got done. And now, with AI involved, they’re suddenly visible. Because if you can’t explain what you want, you won’t get it. If you can’t spot when something’s off, it won’t fix itself. If you don’t apply judgment, the AI won’t apply it for you.

These moments happen subtly and often quickly. Which is why I’ve started thinking of them as invisible differentiators.

So what can you do with this? If you’re working with AI today, or thinking about where to start, here are a few things worth practising:
1️⃣ Say the task aloud before prompting: If you can’t explain it clearly, you’re not ready to brief an AI tool
2️⃣ Re-read outputs with tone in mind: Does this sound like you? Or just correct?
3️⃣ Track your friction points: Where do you instinctively rewrite, reframe, or pull back? That’s where your real skill is showing up.

AI isn’t removing the need for these skills. It’s revealing how much they were doing all along.

What are you noticing in your own work? Are there any skills that used to feel invisible, but now feel essential?

It’s easy to assume that AI is reducing the need for human judgment in legal work. But in practice, I'm finding that it does the opposite, or indeed should be doing so. Once the model gives you something, the real thinking begins.

For instance, you still have to ask:
🔹 Is this right for this client, this tone, this context?
🔹 Does it reflect expertise or just efficiency?
🔹 What’s missing that only a human would notice?

These aren’t editing tasks. They’re judgment calls.

And the more AI becomes part of legal workflows, the more visible those calls become. You stop accepting outputs at face value. You start scanning for structure, logic, nuance, tone. You begin to recognise the telltale signs of what I've previously called the AI accent: the neatness, the balance, the subtle lack of conviction.

You may even catch it in your own writing, especially when you’re short on time or tempted to move quickly. (I’ve seen it happen - and not just to others.)

So what does the AI accent actually sound like? Here are a few things to watch for:
🔹 An overly balanced tone: everything is neatly considered, but nothing stands firmly.
🔹 Fluency without substance: the language flows, but doesn’t say anything new.
🔹 The lack of specificity: generalities in place of real-world context or nuance.
🔹 A tone mismatch: formal when it should be human, or neutral when it should be confident.

These are subtle cues. But they matter. Because AI can get you to a draft. Only judgment gets you to something meaningful. The best legal professionals I’ve seen using AI aren’t the ones getting it to do everything. They’re the ones who know when to step in and what to change.

Have you ever spent more than 30 seconds thinking about the context window when using AI? I have, and I think the time spent was fruitful.

We often treat it as a command box. Just another interface. But I’ve come to see it differently.

It’s a mirror. Not because it reflects you, but because it reflects your thinking.

When you work with a large language model, the quality of its output depends entirely on the clarity, coherence, and completeness of what you give it. And that starts with your own understanding of the task.

In legal work, we pride ourselves on precision. But even experienced professionals can struggle to express what they actually want. Especially when the recipient is no longer a junior lawyer, but a system with no assumptions, no judgment, and no shared context.

That’s when the mirror kicks in.

If you’ve ever struggled to phrase a prompt, it’s not because you don’t know how to use AI. It’s because you’re still figuring out what you actually think.

It doesn’t just ask, “What do you want?”
It quietly asks:
Do you actually understand the structure of the thing you're asking for?

Can you identify the key issues and explain them simply? Do you know what good looks like in terms of tone, length, format, and depth? Or are you hoping the AI will guess?

Poor prompting isn’t always a skills gap. Sometimes, it’s a thinking gap. In this way, AI becomes more than a tool. It becomes a clarity exercise. A way to stress test your brief, your priorities, even your own logic.
So if the context window is a mirror, how can you make better use of it? Here are three things I find helpful:
1️⃣ Say it aloud before you write it. If you can’t explain what you want in one sentence, the AI probably won’t be able to either.
2️⃣ Break the task into parts. Structure, tone, purpose - prompting works best when you separate the pieces.
3️⃣ Write as if you're briefing a human who knows nothing. Because that’s exactly what you’re doing.

As a Legal Engineer, I find this fascinating. We’re not just teaching lawyers how to prompt. We’re watching them sharpen their thinking in the process. Because writing to the model forces them to write for themselves.

The context window won’t tell you what you meant. It reflects what you actually said. Or didn’t. It’s not just a prompt box. It’s an invitation to think more clearly.

Where has AI helped you uncover gaps in clarity or thinking?

Everyone says AI will save you time.
But what if saving time was never the real point?

A lot of the conversation around legal AI focuses on time savings. Summarise faster. Draft quicker. Review in minutes, not hours. And while that’s true, it’s only half the story. AI doesn’t solely save time. It reallocates it. It 𝘵𝘳𝘢𝘥𝘦𝘴 routine for reflection.

Consider the following situations:
🔷 The time you used to spend drafting from scratch? Now you spend refining and sharpening.
🔷 The time you saved reviewing documents? Now you invest in thinking about the bigger picture.
🔷 The minutes you used to spend structuring emails or updates? Now you use them to be clearer, more persuasive, more thoughtful.

From my recent conversations, the lawyers who get the most from AI aren’t the ones looking to "get work done faster." They’re the ones looking to spend their time differently. More on judgment. More on creativity. More on client strategy.

The shift isn't just about efficiency. It's about elevation. When AI takes care of the mechanical layers of legal work, it opens up room for deeper thinking, sharper communication, and more strategic influence (among other things). The real advantage isn’t doing the same work faster. It’s doing better work - work that reflects expertise, not just output.

I'd argue that AI doesn’t take away the hard parts of legal work. It gives you the chance to move through the mechanical parts quicker so that you spend more energy where it really counts. If we only measure AI’s value by "hours saved," we may be missing the point.

The real win is how it reshapes the way we work, think, and create.

Where has AI helped you shift your focus to the work that matters more?

I think there may be a quiet misconception in legal circles: that AI has to deliver a big moment to be valuable.

The “aha!” moment.
The cost / time saved.
The one perfect use case that justifies it all.

And when that moment doesn’t happen, it’s easy to assume AI isn’t working, or worse, that it’s overhyped.

But in my experience, AI’s value doesn’t usually land with fireworks. It shows up as friction reduction. For instance:
🔷 That clause you rewrote faster than usual
🔷 The briefing note you structured in minutes, not hours
🔷 The internal update you polished over lunch instead of after hours
🔷 The five versions of something you didn’t have to write from scratch
🔷 The slides you designed with AI's creative input, turning complex legal ideas into engaging visuals

Nothing dramatic. Just less resistance. Fewer blockers. A little more ease in the work. That’s the shift. And it compounds. The myth of the 'AI moment' makes us wait for something extraordinary. But the real transformation happens in the margins - in how we think, how we write, how we make decisions. Quietly. Iteratively.

If you’re still waiting for AI to blow your mind, maybe take another look. It might already be making your work smoother, just without asking for applause.

Curious: where have you seen AI reduce friction in your day-to-day?

There seems to be a fear among some lawyers I speak to that AI will flatten legal writing. That everything will start to sound the same. That nuance, tone, and voice will disappear in favour of bland, balanced, “AI-sounding” content.

But lately, I’ve been noticing something different.

When used intentionally, not as a ghostwriter, but as a sounding board or redrafting tool, AI can actually help lawyers become more expressive, not less.

Why? Because AI gives you the space to explore style, not just produce content.

For instance:
🔷 It helps you get unstuck faster
🔷 It lets you test different phrasings and tones in seconds
🔷 It turns vague thoughts into clearer drafts you can refine
🔷 And it gives you more time to focus on how something is said, not just what is said

In a world where deadlines are tight and the craft of writing often gets sacrificed in the name of efficiency, AI can bring that craft back into the process.

I’ve seen lawyers use AI to:
1️⃣ Rewrite a long-winded clause to make it more readable
2️⃣ Adjust tone for different stakeholders - more warmth here, more formality there
3️⃣ Simplify dense legal points for non-legal audiences
4️⃣ Try multiple versions of an argument before landing on the strongest one

None of this replaces legal skill or judgment. It supports it. The lawyer is still in the driver’s seat, but now they’re backed by a collaborator that helps them move faster through the messy middle of writing, without losing control of the end result.

In that sense, AI doesn’t dilute personal style. It helps lawyers express it more clearly.

Has AI changed the way you write, communicate, or refine ideas? Has it helped you sound more like yourself?

Ever read something and just know it was written by AI? Even when it’s factually correct and grammatically flawless, there’s something about it - the phrasing, the structure, the overly neutral tone - that gives it away.

That’s the AI accent.

It’s not a single trait but a collection of subtle tells:

🔹 Overuse of phrases like "In today’s fast-paced world..."
🔹 An unnatural balance - every argument neatly weighed, every nuance accounted for, but without real conviction
🔹 A lack of true human fingerprints - no distinctive voice, no risk-taking, no rough edges

For legal professionals, this matters. Legal teams operate in a field where credibility, judgment, and persuasion are paramount. If your writing sounds like AI, it might signal a lack of depth, a lack of personal expertise, or worse, a lack of effort.

The challenge isn’t just spotting the AI accent. It’s learning how to edit it out.

As a Legal Engineer, I see this all the time - lawyers experimenting with AI tools but struggling to get outputs that sound human. My focus is on helping to bridge that gap: understanding why AI writes the way it does and how to fine-tune prompts, edit responses, and use AI as a true assistant rather than a ghostwriter.

Have you noticed the AI accent creeping into legal work? And how do you fix it?

Have you ever used an AI tool and thought, “That sounds completely reasonable” until something didn’t quite sit right?

That instinct matters. And it might be one of the most important things we bring to the table when working with AI.

Last week, OpenAI shared new research into why large language models hallucinate. The finding? These models aren't just struggling with facts. They're shaped by incentives that reward confident, plausible-sounding answers rather than admissions of uncertainty. You can read their paper here: https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4aaa5/why-language-models-hallucinate.pdf.

That has serious implications for legal work. Because when the output looks polished, it becomes easier to miss where the assumptions begin. AI isn’t just making things up. It’s reflecting back a world shaped by the structure of your prompt, the gaps in the data, and the way the model was trained to prioritise sounding right over being right.

I’ve tended to treat hallucinations as a prompt design signal, and that’s helped quite a bit. If something looks too neat, I ask myself: where has this model smoothed over a rough edge? If something sounds too confident, I ask: what does it assume I won’t challenge?

If you’re experimenting with AI in legal work, here are a few approaches that can help sharpen how you use output:

1️⃣ Ask for uncertainty
Prompt the model to flag where it’s unsure or where its answer may depend on missing information.

2️⃣ Test the edges
Explore how the output changes when you vary the jurisdiction, client type, or risk appetite.

3️⃣ Interrogate the frame
After getting an answer, ask the model to critique its own logic or defend the assumptions it relied on.

4️⃣ Stress test with scenarios
Use best case, worst case, and most likely case framing to reveal how the model weighs different outcomes.

5️⃣ Check what the model needs
Ask what information would improve the output or reduce ambiguity, then decide if that information is available.

AI won’t warn you when it's hallucinating. It will present output with the same polish and tone either way. That means your job isn't just to read the output, but to understand the shape of the thinking behind it.

Where have you seen AI sound confident when it shouldn’t have? And how are you building checks into your own prompting?

We spend a lot of time talking about how to write better prompts. But lately, I’ve been thinking more about what happens before you even write one.

Because the prompt is rarely the beginning. The real starting point is how you’re thinking about the task itself: the questions you’re asking, the scope you’re setting, and the shape of the answer you’re expecting.

Before you type anything, you’ve probably already made a dozen decisions:
🔹 What’s the purpose of this task?
🔹 What do I already assume to be true?
🔹 What does “useful” even look like here?

In that sense, your pre-prompt thinking becomes part of the invisible instruction set. And if that framing is misaligned, too narrow, too vague, or too biased, then even the most elegantly written prompt will fall short.

I’ve seen this show up a lot in legal tasks:
🔹 You want a risk summary but haven’t decided what kind of risk matters
🔹 You ask for a clause comparison but haven’t clarified what “better” means
🔹 You prompt for next steps but haven’t defined what stage you’re actually at

So instead of just writing better prompts, I’ve started spending more time thinking around the task first. Framing the real question before prompting the answer.

Here are a few things I’ve found helpful to try:

1️⃣ Scope the decision, not just the task
Ask yourself what needs to be decided or understood, not just what output you want.

2️⃣ Name your assumptions
Before prompting, jot down what you’re already assuming about the situation or context. Then test those.

3️⃣ Clarify success
What would a great answer look like? Not perfect, just genuinely useful. That can shape the tone and structure of your prompt.

4️⃣ Write the summary first
Write the headline of what you hope to get out of the task. Then reverse-engineer the prompt from there.

5️⃣ Try zero-prompting
Before typing anything, paste the source material and ask the model what it thinks the task might be. The gap between your intention and its guess can be illuminating.

Prompts don’t live in isolation. They’re shaped by the thoughts and decisions that come before them. And the more aware you are of those invisible steps, the more powerful your prompting becomes.

What do you do before writing a prompt?

When we talk about prompting, we often focus on what to ask. But what about how we ask it? Not just the structure of the prompt, but the temperature of the thinking we’re inviting.

In technical terms, model temperature is a setting that controls how predictable or creative the AI's response will be. A low temperature keeps things focused and consistent. A high temperature lets the model take more creative or lateral paths. But beyond the setting itself, there’s also the tone and framing of the prompt - how tightly you define the task, how much ambiguity you allow, and how much interpretation you’re inviting.

That’s where I’ve been finding value lately. The real shift comes when the thinking temperature of the prompt fits the job at hand.

🔷 Early-stage tasks often benefit from warmer, looser prompts: You’re not looking for precision, you’re looking for patterns, possibilities, or inspiration.
🔷 Review or validation tasks tend to perform better with colder, more structured prompts: You want clarity, discipline, and fewer surprises.

It’s made me think that prompting isn’t just about precision. It’s about calibration. And knowing when to turn the heat up or down is part of the craft.

If you want to experiment with this idea, here are a few things to try:

1️⃣ Start hot, then cool down
Use a looser prompt to generate ideas or spot angles you might miss. Then follow it up with a more focused version to refine and assess.

2️⃣ Vary your language
Phrasing matters. “List ten creative possibilities…” yields something very different from “Identify the most likely outcome…” even if the task is the same.

3️⃣ Use opposing prompts side by side
Ask the model to approach the same problem from optimistic and pessimistic perspectives. Compare the logic and look for blind spots.

4️⃣ Match temperature to context
Regulatory, policy, and compliance work usually benefits from colder, more rigid prompting. Strategic, design-led, or exploratory tasks tend to improve with warmer thinking.

5️⃣ Use temperature as a diagnostic
If the output feels vague, try cooling it down. If it feels too flat or narrow, consider warming it up.

Prompting is never just about getting the right answer. It's about setting up the right kind of thinking for the task in front of you.

Where have you found it helpful to turn the heat up or down?

Have you ever thought about the half life of prompts? In science, half life refers to the time it takes for something to lose half of its effectiveness. I think a similar idea can apply to AI prompts. It’s not that knowledge itself decays, but the way knowledge is put into use through prompts can. A prompt that feels sharp and effective today might start to feel less precise tomorrow, not because the words have changed, but because the context around them has.

I’ve seen this a lot in legal and compliance work lately. A review prompt that no longer reflects a client’s wants and needs. A due diligence summary that doesn’t quite capture a new regulatory nuance. A drafting instruction tuned for one model, but less effective on the next. The output still looks fine, but the fit isn’t quite right.

What does this mean for prompt engineering? To me it highlights that prompt design isn’t just about the immediate result. It’s also about how resilient a prompt is over time. Engineering a prompt well is still hugely valuable, but we should also expect drift. That means testing, refreshing, and sometimes rebuilding the frame, rather than treating prompts as static instructions.

And what about prompt libraries? I think the real long term value may not just be in storing the prompts themselves, but in capturing the scenarios they represent. Even if a prompt changes, the hard work that went into identifying meaningful use cases and shaping them into something structured doesn’t necessarily lose relevance. Over time, prompt libraries could evolve into scenario libraries that preserve legal and commercial know-how, even if the precise wording needs to be updated.

If you want to explore this idea, here are some things I’m going to start trying:
1️⃣ Anchor to the scenario - record what the prompt is designed to achieve and when it applies.
2️⃣ Add version details - track the model, date, and audience so drift becomes visible.
3️⃣ Test for drift - run the same prompt on a fixed set of materials and compare changes.
4️⃣ Surface assumptions - note what inputs or context the prompt relies on.
5️⃣ Treat libraries as living - update wording, but preserve the scenario and reasoning that gave rise to it.

Have your prompts started showing their half life?

In case you missed it, OpenAI released GPT-5 last week. I’ve been testing it over the past few days, and one thing that struck me is how faithfully AI reflects what you put in. It sounds obvious, but in practice it can sometimes be more subtle. You give the model a prompt, it works within the frame you set, and the result often feels clean and confident.

The trouble is, if your starting point’s flawed, the model will carry that flaw all the way through. It won’t stop to ask whether the question was the right one, whether the scope was too narrow, or whether you left something important unsaid.

In the absence of a suitable technical term, I'm starting to think of this as the 'AI echo'. The model mirrors your assumptions so closely that it can be hard to see where those assumptions began.

I’ve seen it most often in legal work that demands nuance – contract review, due diligence, or drafting with a specific commercial lens. If you don’t test the frame, the model will keep it intact, even when it’s misaligned.

That’s why I treat part of my prompting as an exercise in disruption. I’ll flip the scenario, switch the audience, or ask the model to tell me what I might be missing. Sometimes I’ll strip away my structure completely to see how it approaches the task from scratch.

If you want to avoid the AI echo, here are a few approaches worth trying:

1️⃣ Flip the frame - Ask the model to tackle the task from the opposite perspective or under different constraints.
2️⃣ Surface the missing - Prompt it to identify the information or context it would need to improve the output.
3️⃣ Test extremes - Explore how it responds if you dramatically over- or under-scope the task.
4️⃣ Strip it back - Remove your structure entirely and see what it produces unprompted.
5️⃣ Challenge the safe answer - Ask it to defend or critique its own reasoning rather than just present it.

AI won’t challenge your framing by default. But you can design for that. And often, the most valuable output is the one that breaks the echo.

Have you picked up on the AI echo?

AI often gets talked about as a replacement. For time. For effort. For human thinking.

But in practice, especially in legal work, that’s rarely the real story. The value of AI isn’t always in what it takes away. Sometimes, it’s in what it gives back. One of the most powerful things it offers is perspective.

I’ve been working through a particularly complex governance process document recently. A dense, multi-layered text with logic that branches and loops. The kind of task that resists shortcuts. It’s not just about identifying steps. It’s about understanding what drives them, how they connect, and where decisions hinge on hidden rules.

AI didn’t give me the answer. It didn’t automate the work. What it did was shift how I approached it. It surfaced structure I hadn’t yet seen. It gave me a starting point I could react to. It let me zoom out before zooming in, giving shape to something that initially felt shapeless.

That didn’t make the work easier. But it made the thinking behind the work sharper.

The output itself wasn’t the end goal. It was a way to see what mattered. To notice assumptions. To challenge framing. And in that sense, AI didn’t replace judgment. It supported it. The more complex the task, the more valuable that support becomes.

If you’re using AI to work through complexity - a process, a policy, a system - here are a few prompts I’ve found helpful:

1️⃣ Where is the tool helping me see structure, not just produce content?
2️⃣ What assumptions is it surfacing that I might have missed?
3️⃣ How can I use this output to refine or challenge how I’m thinking?

AI can’t always resolve ambiguity. But it can help you move through it with more clarity and control.

Where has AI helped you step back, notice something new, or reframe a task you thought you understood?

There’s still a common (and understandable) belief that AI works like a magic box: type in a prompt, and out comes a polished result. When the output misses the mark, it’s easy to assume the tool isn’t good enough or that AI just doesn’t “get it.”

But often, the issue isn’t the AI itself. It’s 𝘩𝘰𝘸 it’s being briefed.

One of the most useful mindset shifts I’ve seen is to stop thinking of AI as something you “ask” once and instead treat it more like a capable collaborator that benefits from clear, structured direction.

This collaborator:
🔷Performs best when it knows what’s expected.
🔷Needs context to provide relevant answers.
🔷Works more effectively when you break down complex tasks into manageable steps.
🔷Will often produce something vague or overly cautious if you ask for too much at once.

Of course, time is a real constraint. For most lawyers, the idea of spending extra time to craft better prompts or iterate on outputs can feel like yet another thing to fit into an already full day.

But skipping that setup usually leads to more frustration later - poor outputs, rewrites, and extra rounds of clarification. Spending just a few more minutes to clarify what you need, how you want it structured, and what the goal is can dramatically improve results. And if the task is repeatable, that effort compounds.

In my work as a Legal Engineer, I’ve found that some of the biggest improvements don’t come from new tools, but from better use of existing ones. Prompting well isn’t about fancy language. It’s about understanding the process behind the task and guiding the AI through that logic.

When legal teams treat AI as a sparring partner, one that works best with structure and intent, they see stronger, faster results. AI can be an incredibly effective collaborator. But it still needs direction.

How are you learning to brief AI in a way that works for you?

I’ve been having a lot of conversations lately with people frustrated by AI tools that aren’t giving them great results. Nine times out of ten? They’re using one monolithic prompt for a task that’s actually pretty complex.

This is what I’ve started calling prompt debt - the cost of trying to do too much with a single prompt, only to spend more time correcting, regenerating, or abandoning the output altogether.

The fix isn’t better wording. It’s better thinking.

When a human does the task, they don’t just launch into it. They think, ask questions, check documents, build structure, refine tone. Legal work is inherently multi-step. So why do we expect AI to perform like a senior associate off the back of one vague sentence or a jumbled wall of text?

A better approach is what’s known as prompt chaining - breaking the task down into steps, stages, or checkpoints. Often, this starts by evaluating the complexity of the task itself:
🔷 What would a human need to do this well?
🔷 Where are the points of discretion, subjectivity, or risk?
🔷 How do we simplify or sequence that for AI to follow?

As Legal Engineers, Hannah Beddard and I would say that this is where process mapping becomes unexpectedly useful. Prompting well isn’t just about knowing the right syntax. It’s about understanding the shape of the legal process behind the task, and approximating that shape in AI form.

AI isn’t a magic box you throw instructions at. It’s a collaborator that works best when you guide it - not with one big ask, but with a structured flow.

Have you noticed prompt debt creeping in? And what techniques are helping you get better outputs faster?

Legal AI is having a moment. Everywhere you look, there’s a new list of use cases: summarise this, review that, draft it faster. And that’s excellent - AI is helping legal teams work more efficiently, and in some cases, rethink how work gets done altogether. There couldn't be a more fun time to be a Legal Engineer.

But in all the excitement (and following on from my post last week), I think we’re missing an equally valuable part of the conversation: when AI isn't the right fit.

Not because the tools aren’t good enough. Not because lawyers are resistant. But because sometimes, the best decision may not be to automate, delegate, or summarise.

That might be:
🔷 When the stakes are high and tone matters more than turnaround
🔷 When the data going in is unstructured, messy, or sensitive, and human review is just safer
🔷 When the bottleneck isn’t actually a drafting or review task, but something more structural or relational
🔷 Or when AI would be a shiny distraction from what’s really going on

And in those moments, the answer might not be AI at all. It could be:
✅ A rule-based or deterministic solution that’s easier to govern and more predictable
✅ A simple automation that removes repetitive manual steps
✅ A better-designed workflow that avoids the problem entirely
✅ Or a fresh perspective using legal design to rethink how the work gets done

Knowing when not to use AI isn’t about being cautious - it’s a sign of strategic maturity. It’s about treating AI like what it is: a powerful tool, but not always the right one. In my work as a Legal Engineer, I suspect I’ll spend just as much time helping teams spot the right moments for AI as I will helping them confidently say, “Actually, this one’s better solved another way” in the future.

We often celebrate the clever use cases, and rightly so. But sometimes, the most impactful thing you can do is recognise an anti–use case and decide to leave it alone.

Would love to hear from others: have you ever chosen not to use AI for something, and felt it was the right call?

Legal AI is often framed around what it can do - summarisation, analysis, review etc. But lawyers don’t work in isolated tasks. They work in processes: closing deals, responding to regulatory inquiries, preparing for trial. AI is most valuable when it’s not just performing a function but actually fitting into the way legal work happens: streamlining workflows, reducing bottlenecks, and letting lawyers focus on what matters.

This is where legal engineering comes in. It’s not just about introducing AI tools, it’s about making sure they actually work for lawyers. And that starts with asking the right questions.

🔹 Where are the biggest pain points?
AI can’t just be dropped into a process for the sake of it. The starting point has to be what’s slowing lawyers down, not what AI can do.

🔹 How do we translate legal work into AI solutions?
Lawyers don’t think in terms of “prompt engineering” or “workflow automation.” They think about getting deals over the line, responding to clients, managing risk. Legal engineers bridge that gap, turning legal problems into something AI can actually solve.

🔹 Where does AI fit into the flow of work?
AI adoption fails when it asks lawyers to change how they work. The best implementations don’t add extra steps, they remove friction. That could mean pre-extracting key clauses before contract negotiation, flagging regulatory updates for compliance teams, or helping litigation teams pull together chronologies faster.

The real challenge isn’t figuring out what AI can do. It’s figuring out where AI makes a real difference. It's ensuring AI doesn’t just exist in theory, but actually integrates into the way lawyers already work.

Where do you see AI fitting into your workflow?

A week ago, I wrote about the AI accent, the subtle but unmistakable way AI-generated text feels different from human writing. The response to that post confirmed what I suspected: legal professionals are noticing it, and many may be wondering what to do about it.

So, let’s talk solutions.

But first, where does the AI accent come from?
🔹 Prediction, not intention. AI doesn’t think - it predicts. Every word it generates is based on statistical likelihood rather than deep understanding. This is why it tends to hedge, weigh every argument equally, and avoid strong opinions. It optimises for plausibility rather than conviction.
🔹 Fluency over substance. AI is designed to sound polished. But fluent text can mask shallow reasoning or factual gaps, leading to output that feels correct but lacks meaningful insights.
🔹 Data-driven blandness. AI is trained on massive datasets, so it defaults to the most common, widely accepted ways of writing. The result is neutral, general, and risk-averse content. Good for avoiding controversy, bad for sounding distinctive.

How do we manage this in legal work?

1️⃣ Expectation-Setting: AI is a Co-Pilot, Not an Author
One of the biggest mistakes I see is lawyers expecting AI to produce polished, client-ready documents. The reality is different. AI is a starting point, not a finished product. Treat AI-generated text like a first draft from a junior colleague. It needs review, refinement, and a human touch, but ultimately it's accelerating the drafting process.

2️⃣ Better Inputs, Better Outputs: Prompt Engineering for Precision
If you start with vague prompts, you’ll get vague responses. The more structure and specificity you provide, the more useful and human-like the output becomes.

❌ Instead of asking:
“Summarise the key provisions of the EU AI Act.”

✅ Try:
“Summarise the key provisions of the EU AI Act related to high-risk AI systems in plain English. Keep it under 150 words and focus on key obligations and points of interest for our client, [insert brief description of client].”

3️⃣ Layering Human Judgement: Editing AI for Depth and Voice
Even with better prompts, AI’s writing often lacks the persuasive nuance that legal work requires. This is where human-in-the-loop editing is essential.

Use AI for structure and speed, but refine the language to add:
🔹 Credibility – Does the response reflect expertise and sound authoritative?
🔹 Persuasion – Does it read like an argument, not just an information dump?
🔹 Personality – Does it sound like you, or just generic legalese?

The AI accent may never go away, so managing it is an essential skill in legal work. AI should help lawyers work faster, not dilute their work.

Reading for pleasure is such a powerful thing. You don’t always realise what sticks - until it surfaces years later, reshaped into something useful. That’s what’s inspired this week’s reflection.

I’ve always treated AI as a tool, something that I instruct and guide. And I think that’s still my philosophical view of the tech. However, the legal engineer in me has started to wonder whether it’s also a kind of second brain for legal professionals.

Not in the sci-fi sense. More in the Tiago Forte sense: an externalised system that reflects how you think, make decisions, and structure your work.

It starts small. A saved prompt here. A redraft you liked. A structure that just works.

But over time, something shifts. You’re no longer using AI as a one-off assistant. You’re creating workflows that extend your legal thinking by codifying your logic, tone, and instinctive process.

For example:
🔹 You start developing prompt templates that mirror how you summarise, rewrite, or redline
🔹 You create reusable scaffolds that help guide analysis or decision-making
🔹 You review outputs so consistently that your edits become patterns, and those patterns get built into how you prompt next time

Eventually, you’ve got something more than a chatbot. You’ve got a system. One that knows what good looks like for you. One that reduces repetition, preserves consistency, and frees you up to focus on the decisions that matter.

That’s what I think AI is becoming in legal. Not a replacement for thinking. A structure that holds it. In this way, it starts to act like a legal second brain. A companion that:
🧠 Remembers your approach to tasks
🧠 Mirrors how you balance tone, risk, and nuance
🧠 Keeps getting better the more you interact with it
🧠 Reflects your judgment, not replaces it

It’s also a subtle shift in where value lives. Not in a single output. In how those outputs compound.

The real gains come when you stop treating prompts as one-offs and start treating them as assets. When you stop thinking of AI as a typing assistant and start seeing it as an extension of your legal reasoning. A living system you shape over time.

So if you’re using AI day to day, here’s a few questions I’ve been asking myself:
🔹 What am I training this system to remember?
🔹 What patterns am I feeding it?
🔹 What kind of second brain am I building, and how intentional am I being about it?

Would love to hear from others: What’s showing up in your own second brain?

Last year, Drew Winlaw gifted me a signed copy of Madhumita Murgia’s 'Code Dependent: Living in the Shadow of AI'. One of its biggest takeaways? AI doesn’t just exist, it reshapes the systems we build around it, for better or worse. It can expose inefficiencies and structural flaws, but if we don’t address the root problem, it often reinforces them instead.

That got me thinking about legal work. There’s a temptation to see AI as a quick fix. A process is slow? Automate it. Reviewing documents takes too long? Use AI to summarise. But if the process itself is broken, AI won’t solve the issue. It'll just make a bad system run faster.

That doesn’t mean AI isn’t valuable. Its real power isn’t just in automating existing workflows but in forcing legal teams to ask why those workflows exist in the first place. AI can be a catalyst - a reason to step back and rethink how legal work is structured.

🔹 If contract review takes too long, is AI really the answer? Or is the real issue that risk thresholds aren’t clear, leading to unnecessary escalations?
🔹 If compliance reporting is overwhelming, should AI be automating summaries? Or should teams be rethinking how obligations are tracked in the first place?
🔹 If a deal involves endless email back-and-forth, is AI-powered drafting the best fix? Or is there an opportunity to redesign approval flows entirely?

This is where I think legal design comes in. Instead of using AI to patch over existing pain points, legal teams should be asking, What's the best way to achieve this outcome? AI then becomes part of a larger shift, not just making old processes faster but enabling entirely new ways of working.

AI isn’t just a tool for automation. It’s a tool for transformation. But only if we use it to challenge, not just accelerate, the status quo.

Where do you see AI pushing legal teams to rethink how they work?

In case you missed it, Vals AI released their Legal AI Report last week, a truly gripping read I thought. The report highlighted a shift that’s hard to ignore - AI tools are no longer just supporting legal work, they’re matching or even surpassing human lawyers on tasks like document summarisation and data extraction. While this confirms AI’s potential, it also raises a deeper question: if AI can handle core legal tasks with increasing accuracy, what does that mean for the role of the lawyer? And more personally, does using AI feel like cheating?

For years, skill in law was defined by how well you could write, analyse, and think - not how well you could prompt a machine. When AI takes on those cognitive tasks, it can feel like skipping the hard part or outsourcing expertise. But is that feeling justified?

When AI feels like cheating (and when it doesn’t):
🔹 Relying vs delegating. There’s a difference between using AI to enhance thinking and using it to replace thinking. If AI is doing all the work, are we still the expert?
🔹 The Junior Associate test. If a lawyer handed off a task to a junior, would that be “cheating”? Probably not. So why does it feel different when AI is involved?
🔹 Mental effort as a badge of honour. There’s an ingrained idea that the harder something is, the more valuable it must be. AI disrupts that by making some tasks effortless. But does ease reduce value, or just redefine it?

So, how should we think about AI in legal work?
✅ AI isn’t a shortcut to skill. It’s a shortcut to output.
The value isn’t in doing everything manually but in applying judgment to AI-assisted work.
✅ Knowing how to use AI well is a skill in itself.
Lawyers don’t feel guilty for using templates or precedent banks. AI is just a more advanced tool.
✅ Expertise isn’t disappearing, it’s shifting.
The best professionals won’t be the ones who avoid AI but the ones who know when to trust it, when to challenge it, and when to override it.

The Vals AI report highlights that AI is already an integral part of legal workflows. We don’t question spellcheck or search engines anymore, even though they do work that once required human effort. AI is just the next evolution. So does it really feel like cheating, or is that just an old instinct we haven’t shaken yet?

Curious to hear what others think. Have you ever felt uneasy about using AI?

We spend a lot of time talking about what AI can learn. But not enough about whether it can forget.

Machine unlearning is the idea that a system should be able to forget something it previously learned, whether that’s due to a data deletion request, a regulatory obligation, or an error in the training set. As more organisations embed models into day-to-day decision-making, this is becoming less of a fringe technical curiosity and more of a pressing legal and operational concern.

The technical questions are important.

🔷 How do you remove a datapoint from a model without retraining the entire thing?
🔷 What kinds of memory are actually being cleared?
🔷 What do you do about outputs that were generated based on information that is no longer valid?

But the design and governance questions are just as interesting. From a legal engineering perspective, I find myself asking different things. For instance:

🔷 What kind of operational memory does a system have?
🔷 If an AI output informed a contract clause or triggered a decision, can you trace that? Can you reverse it?
🔷 What does accountability look like if we don't just need to prove what the model knew, but also when it knew it?

These are big questions, but also very real ones. Especially for internal-facing systems and legal automation tools. And with growing regulatory focus on explainability, traceability, and the right to erasure, these questions are only becoming more urgent. If you’re advising on or building systems that use AI, here are a few areas worth exploring:

1️⃣ Traceability
Can you trace which prompt, source, or dataset contributed to a particular output? Is that visible to users?

2️⃣ Designing for decay
Could your system include expiry dates for certain types of knowledge? Could it signal when a prompt or dataset is due for review?

3️⃣ Prompt accountability
Are prompts in your libraries documenting their assumptions and context? Could those prompts outlive their usefulness?

4️⃣ Retrospective cleanup
If something is unlearned or retracted, how do you identify the downstream work it influenced? What workflows help with this?

5️⃣ System roles and responsibilities
Who decides what gets forgotten? Who documents it? And how do those roles get built into the system?

This week’s post isn’t necessarily a how-to. It’s more of a what-if. But I think these what-ifs are the kinds of questions lawyers, engineers, and governance teams need to ask together. It’s something our lawyers and technologists at Simmons & Simmons are actively helping organisations navigate.

Curious to hear from others: have you come across machine unlearning before? If so, what practical trade-offs do you think we’ll need to get more comfortable with once techniques mature?

I’ve been thinking about privacy by design lately. Not just as a legal principle, but as a mindset and engineering reflex. One that reshapes how, when, and why legal teams engage with product decisions.

If you haven’t come across it before, ‘privacy by design’ means building privacy into an AI system from the very start. Not retrofitting it once the product’s live. It’s the idea that privacy should influence how personal data is collected, handled, stored, and accessed - not just how it’s described in a policy.

And in practice, that means asking better questions, sooner. For instance:

🔷 What data set is being collected, and is it truly necessary?
🔷 How transparent is the system and how informed is the data subject?
🔷 What are the safeguards around edge cases and misuse?
🔷 Who’s accountable when different privacy risks materialise?
🔷 Are privacy controls designed for real-world use and not just policy and theory?

These aren’t hypotheticals. They’re design decisions with legal consequences. When privacy becomes a design problem, lawyers need a different mindset. They’re collaborating with product teams to shape what actually gets built and how it gets deployed.

That’s where I think legal engineering really comes in. Not to give privacy law advice, but to operationalise it. My role sits in the messy middle: helping technology and legal teams understand each other, translating principles like privacy by design into process logic, workflow, and system design. It’s about identifying the right questions early, designing guardrails that make sense in practice, and making sure legal thinking actually shows up in the build, not just the policy.

And that’s why I’ve loved working with our Middle East Digital Business team at Simmons & Simmons (led by the brilliant Raza Rizvi) on some mind-blowing and pioneering AI projects for TMT and FI sector clients. It’s where legal, product, regulatory, and privacy thinking collide, and where multidisciplinary collaboration is really delivering value. Our firm’s sector focus enables us to get much deeper into our clients’ products and services, meaning that principles like privacy by design aren’t just abstract ideals.

Are there particular principles you think should guide how we design AI? Let’s discuss what you’re doing to bring these principles to life.

Earlier this year, I read Mel Robbins’ excellent book “The Let Them Theory.” The idea articulated is relatively simple: when people behave in ways you can’t control, let them. Stop resisting. Reclaim your energy for what you can do instead. Incredibly transformative, and something I’ve gotten tremendous value from. I encourage everyone give it a read (+ check out her podcast).

It struck me recently that this mindset has a surprising relevance to how legal teams are working with AI.

I often see lawyers exert energy trying to make AI do things it’s not suited for. Or worse, they mistrust it entirely and try to do everything themselves. They edit every word, double-check every suggestion, or avoid using the tool until it’s “perfect.” But maybe that’s the wrong approach.

What if effective AI adoption starts with acceptance? What if the key is knowing what to let go of?

Let AI:
🔷 Structure your first draft so you aren’t starting from a blank page
🔷 Summarise routine material so you can focus on interpretation
🔷 Extract clauses or compare terms across documents
🔷 Reframe content for different audiences or stakeholders
🔷 Surface issues for review without needing to decide them

Let lawyers:
🔷 Challenge assumptions that the model might overlook
🔷 Read between the lines and understand nuance
🔷 Spot commercial dynamics or client sensitivities
🔷 Know when to pause or reframe the issue altogether
🔷 Decide what “good” actually looks like in a specific context

Trying to make AI behave like a lawyer creates friction. Trying to act like a machine may just lead to burnout. The best-performing teams I’ve seen are doing neither. Instead, they’re drawing clearer lines. They’re not using AI to replace legal thinking. They’re using it to relocate it.

By letting go of the parts that don’t need your time or judgment, you free yourself up to focus on the parts that do. And over time, you begin to see where your value truly lies - not in producing outputs faster, but in shaping the decisions around them.

👉 Let AI support the process.
👈 Let your expertise shape the outcome.

Where have you learned to draw that line?